Click here to load reader

The Sprite Network Operating voelker/cse221/papers/sprite-ieee-computer88.pdf · PDF file Sprite is similar to LOCUS in that it provides complete transparency, so that the behavior

  • View
    2

  • Download
    0

Embed Size (px)

Text of The Sprite Network Operating voelker/cse221/papers/sprite-ieee-computer88.pdf · PDF file...

  • The Sprite Network Operating System

    John K. Ousterhout Andrew R. Cherenson

    Frederick Douglis Michael N. Nelson

    Brent B. Welch

    Computer Science Division Department of Electrical Engineering and Computer Sciences

    University of California Berkeley, CA 94720

    Abstract

    Sprite is a new operating system for networked uniprocessor and multiprocessor workstations with large physical memories. It implements a set of kernel calls much like those of 4.3 BSD UNIX, with extensions to allow processes on the same workstation to share memory and to allow processes to mi- grate between workstations. The implementation of the Sprite kernel contains several interesting features, including a remote procedure call facility for communication between kernels, the use of prefix tables to implement a single file name space and to provide flexibility in administering the net- work file system, and large variable-size file caches on both client and server machines, which provide high performance even for diskless workstations.

    Keywords: Operating systems, networks, file systems, process migration, remote procedure call.

  • The Sprite Network Operating System November 19, 1987

    1. Introduction

    Sprite is an experimental network operating system under development at the University of Cali-

    fornia at Berkeley. It is part of a larger research project called SPUR (‘‘Symbolic Processing Using

    RISCs’’), whose goal is to design and construct a high-performance multiprocessor workstation with

    special hardware support for Lisp applications [4]. Although one of Sprite’s primary goals is to sup-

    port applications running on SPUR workstations, we hope that the system will work well for a variety

    of high-performance engineering workstations. Sprite is currently being used on Sun-2 and Sun-3

    workstations.

    The motivation for building a new operating system came from three general trends in computer

    technology: networks, large memories, and multiprocessors. In more and more research and engineer-

    ing organizations, computing occurs on personal workstations connected by local-area networks, with

    larger time-shared machines used only for those applications that cannot achieve acceptable perfor-

    mance on workstations. Unfortunately, workstation environments tend to suffer from poor perfor-

    mance and difficulties of sharing and administration, due to the distributed nature of the systems. In

    Sprite, we hope to hide the distribution as much as possible and make available the same ease of shar-

    ing and communication that is possible on time-shared machines.

    The second technology trend that drove the Sprite design is the availability of ever-larger physi-

    cal memories. Today’s engineering workstations typically contain 4 to 32 Mbytes of physical

    memory, and we expect memories of 100-500 Mbytes to be commonplace within a few years. We

    believe that such large memories will change the traditional balance between computation and I/O by

    permitting all of users’ commonly-accessed files to reside in main memory. The ‘‘RAMdisks’’ avail-

    able on many commercial personal computers have already demonstrated this effect on a small scale.

    One of our goals for Sprite is to manage physical memory in a way that maximizes the potential for

    file caching.

    The third driving force behind Sprite is the imminent arrival of multiprocessor workstations.

    Workstations with more than one processor are currently under development in several research organ-

    izations (UCB’s SPUR, DEC’s Firefly, and Xerox’s Dragon are a few prominent examples) and we

    - 1 -

  • The Sprite Network Operating System November 19, 1987

    expect multiprocessor workstations to be available from several major manufacturers within a few

    years. We hope that Sprite will facilitate the development of multiprocessor applications, and that the

    operating system itself will be able to take advantage of multiple processors in providing system ser-

    vices.

    Our overall goal for Sprite is to provide simple and efficient mechanisms that capitalize on the

    three technology trends. This paper is a discussion of the ways that the trends affected the design of

    Sprite. In areas where the technology factors did not suggest special techniques, we modelled the sys-

    tem as closely as possible after Berkeley UNIX.

    The technology trends had only a minor impact on the facilities that Sprite provides to applica-

    tion programs. For the most part, Sprite’s kernel calls are similar to those provided by the 4.3 BSD

    version of the UNIX operating system. However, we added three additional facilities to Sprite in order

    to encourage resource sharing: a transparent network file system; a simple mechanism for sharing

    writable memory between processes on a single workstation; and a mechanism for migrating

    processes between workstations in order to take advantage of idle machines.

    Alhthough the technology trends did not have a large effect on Sprite’s kernel interface, they

    suggested dramatic changes in the kernel implementation, relative to UNIX. This is not surprising,

    since networks, large memories, and multiprocessors were not important issues in the early 1970’s

    when the UNIX kernel was designed. We built the Sprite kernel from scratch, rather than modifying

    an existing UNIX kernel. Some of the interesting features of the kernel implementation are:

    � The kernel contains a remote procedure call (RPC) facility that allows the kernel of each works-

    tation to invoke operations on other workstations. The RPC mechanism is used extensively in

    Sprite to implement other features, such as the network file system and process migration.

    � Although the Sprite file system is implemented as a collection of domains on different server

    machines, it appears to users as a single hierarchy that is shared by all the workstations. Sprite

    uses a simple mechanism called prefix tables to manage the name space; these dynamic struc-

    tures facilitate system administration and reconfiguration.

    - 2 -

  • The Sprite Network Operating System November 19, 1987

    � To achieve high performance in the file system, and also to capitalize on large physical

    memories, Sprite caches file data both on server machines and client machines. A simple cache

    consistency mechanism guarantees that applications running on different workstations always

    use the most up-to-date versions of files, in exactly the same fashion as if the applications were

    executing on a single machine.

    � The virtual memory system uses ordinary files for backing storage; this simplifies the implemen-

    tation, facilitates process migration, and may even improve performance relative to schemes

    based on a special-purpose swap area. Sprite retains the code segments for programs in main

    memory even after the programs complete, in order to allow quick start-up when programs are

    reused. Lastly, the virtual memory system negotiates with the file system over physical memory

    usage, in order to permit the file cache to be as large as possible without degrading virtual

    memory performance.

    � Sprite guarantees that processes behave the same whether migrated or not. This is achieved by

    designating a home machine for each process and forwarding location-dependent kernel calls to

    the process’s home machine.

    The rest of this paper describes the Sprite facilities and implementation in more detail. Section 2

    discusses the unusual features of Sprite’s application interface, and Sections 3-7 provide more details

    on the aspects of the kernel implementation summarized above. Section 8 gives a brief status report

    and conclusions.

    2. Application Interface

    Sprite’s application interface contains little that is new. The Sprite kernel calls are very similar

    to those provided by the Berkeley versions of UNIX (we have ported a large number of traditional

    UNIX applications to Sprite with relatively little effort). This section describes three unusual aspects

    of the application interface, all of which can be summed up in one word: sharing. First, the Sprite file

    system allows all of the disk storage and all of the I/O devices in the network to be shared by all

    processes so that processes need not worry about machine boundaries. Second, the virtual memory

    - 3 -

  • The Sprite Network Operating System November 19, 1987

    mechanism allows physical memory to be shared between processes on the same workstation so that

    they can extract the highest possible performance from multiprocessors. Third, Sprite implements pro-

    cess migration, which allows jobs to be offloaded to idle workstations and thereby allows processing

    power to be shared.

    2.1. The File System

    Almost all modern network file systems, including Sprite’s, have the same ultimate goal: net-

    work transparency. Network transparency means that users should be able to manipulate files in the

    same ways they did under time-sharing on a single machine; the distridbuted nature of the file system

    and the techniques used to access remote files should be invisible to users under normal conditions.

    The LOCUS system was one of the first to make transparency an explicit goal [10]; other file s

Search related